## Describe the bug

When a streaming response contains tool calls (or logprobs), LiteLLM crashes mid-stream with:

``` TypeError: 'MockValSer' object is not an instance of 'SchemaSerializer' ```

The stream terminates early and the client receives no tool call content.

## Root cause

`openai._models.BaseModel` sets `defer_build=True`, so all openai SDK subclasses start with `MockValSer` as their pydantic serializer. The deferred build fires on `__init__`, but the SDK always parses streaming responses via `model_validate()`, which does not trigger the build. So objects like `ChoiceLogprobs` and `ChoiceDeltaToolCall` retain `MockValSer` after parsing.

`streaming_handler.py` stores these raw SDK objects as extra field values in `ModelResponseStream`. When `proxy_server._serialize_streaming_chunk()` calls `model_dump_json()`, pydantic's Rust core looks up `type(value).__pydantic_serializer__` at the C level (no Python `__getattr__` fallback), finds `MockValSer`, and raises `TypeError`.

The crash is intermittent: `model_dump()` accidentally self-heals via `MockValSer.__getattr__` -> `attempt_rebuild()`. If a usage chunk triggers `model_dump()` first, the class is rebuilt as a side e…